emaballarin / fitgabor

[DepsFix 🛠️] A simple tool to find the Gabor filter that maximizes the output of your neuron model

Fit Gabor

If you have a neuron model and you want to find the gabor filter that maximizes the activity of your model neuron, this tool might come handy - it finds the input arguments to a gabor-generating function via gradient descent.

Example

1. We need a neuron model

Let's create a simple linear-nonlinear neuron model with a known receptive field:

import torch
from torch import nn
from torch.nn import functional as F

class Neuron(nn.Module):
    def __init__(self, rf):
        super().__init__()
        self.rf = torch.tensor(rf.astype(np.float32))
        
    def forward(self, x):
        return F.elu((x * self.rf).sum()) + 1


from fitgabor.utils import gabor_fn

theta = -np.pi/4
groundtruth_rf = gabor_fn(theta, sigma=4, Lambda=14, psi=np.pi/2, gamma=1, center=(15, 5), size=(64, 64))
neuron = Neuron(groundtruth_rf)

Here is the ground truth RF:

2. Initialize and train the gabor generator

from fitgabor import GaborGenerator, trainer_fn

gabor_gen = GaborGenerator(image_size=(64, 64))
gabor_gen, _ = trainer_fn(gabor_gen, neuron)

You can generate the learned gabor simply by calling the model:

learned_gabor = gabor_gen().data.numpy()

And here is the learning evolution:

Cite as

@software{Bashiri_fitgabor_2020,
author = {Bashiri, Mohammad},
month = dec,
title = {{fitgabor}},
url = {https://github.com/mohammadbashiri/fitgabor},
version = {0.0},
year = {2020}
}

Tried it and didn't work?

Let me know, please! I would love to know why it did not work and help fix it 👷